IronOCRとDynamsoft OCRの比較

Both sets of software offer solutions for cropping images for OCR.

Reading cropped regions with IronOCR

Iron's branch of Tesseract OCR is adept at reading specific regions of images, as shown in the following code sample.

We use System.Drawing.Rectangle to describe the exact region of an image to be read in pixels.

When dealing with a standardized form that is filled out, and only a portion of the content changes from case to case, this can be really handy.

Scanning a Section of a Page: We use System.Drawing.Rectangle to designate a region to read from a document, improving speed and accuracy.

using IronOcr;
using System.Drawing;

class Program
{
    static void Main(string[] args)
    {
        // Create an instance of IronTesseract
        var ocr = new IronTesseract();

        using (var input = new OcrInput())
        {
            // Define content area of interest
            var contentArea = new Rectangle() { X = 215, Y = 1250, Height = 280, Width = 1335 };

            // Add the specific region to the input
            input.AddImage("img/ComSci.png", contentArea);

            // Perform OCR operation
            var result = ocr.Read(input);

            // Output recognized text to console
            Console.WriteLine(result.Text);
        }
    }
}

using IronOcr;
using System.Drawing;

class Program
{
    static void Main(string[] args)
    {
        // Create an instance of IronTesseract
        var ocr = new IronTesseract();

        using (var input = new OcrInput())
        {
            // Define content area of interest
            var contentArea = new Rectangle() { X = 215, Y = 1250, Height = 280, Width = 1335 };

            // Add the specific region to the input
            input.AddImage("img/ComSci.png", contentArea);

            // Perform OCR operation
            var result = ocr.Read(input);

            // Output recognized text to console
            Console.WriteLine(result.Text);
        }
    }
}

Imports IronOcr
Imports System.Drawing

Friend Class Program
	Shared Sub Main(ByVal args() As String)
		' Create an instance of IronTesseract
		Dim ocr = New IronTesseract()

		Using input = New OcrInput()
			' Define content area of interest
			Dim contentArea = New Rectangle() With {
				.X = 215,
				.Y = 1250,
				.Height = 280,
				.Width = 1335
			}

			' Add the specific region to the input
			input.AddImage("img/ComSci.png", contentArea)

			' Perform OCR operation
			Dim result = ocr.Read(input)

			' Output recognized text to console
			Console.WriteLine(result.Text)
		End Using
	End Sub
End Class

This results in a 41 percent boost in speed, while also allowing us to be more specific. This is extremely valuable for .NET OCR applications involving documents that are comparable and consistent, including invoices, receipts, checks, forms, expense claims, and so on.

When reading PDFs, ContentAreas (OCR cropping) is also supported.

Reading cropped regions with Dynamsoft OCR

To begin, launch Visual Studio and build a new C# Windows Forms Application, or open an existing one. We will need to include DynamicDotNetTWAIN.dll, DynamicOCR.dll, and the appropriate language package.

1. Navigate to Tools -> Choose Toolbox Items, then to the .NET Framework Components tab, click the Browse button, and locate DynamicDotNetTWAIN.dll.

2. Right-click the project file in Solution Explorer and select Add-> Existing Item... then add the necessary items from the OCR resources directory.

This is the code for clicking the LoadImage button:

private void button1_Click(object sender, EventArgs e) 
{
    OpenFileDialog filedlg = new OpenFileDialog(); 
    if (filedlg.ShowDialog() == DialogResult.OK) 
    { 
        dynamicDotNetTwain1.LoadImage(filedlg.FileName);
        // Choose an image from your local disk and load it into Dynamic .NET TWAIN
    } 
}

private void dynamicDotNetTwain1_OnImageAreaSelected(short sImageIndex, int left, int top, int right, int bottom) 
{
    dynamicDotNetTwain1.OCRTessDataPath = "../../"; 
    dynamicDotNetTwain1.OCRLanguage = "eng";
    OcrResultFormat ocrResultFormat = Dynamsoft.DotNet.TWAIN.OCR.ResultFormat.Text;

    byte [] sbytes = dynamicDotNetTwain1.OCR(dynamicDotNetTwain1.CurrentImageIndexInBuffer, left, top, right, bottom);
    // OCR the selected area of the image

    if (sbytes != null) 
    { 
        SaveFileDialog filedlg = new SaveFileDialog(); 
        filedlg.Filter = "Text File(*.txt)| *.txt"; 
        if (filedlg.ShowDialog() == DialogResult.OK) 
        {
            FileStream fs = File.OpenWrite(filedlg.FileName); 
            fs.Write(sbytes, 0, sbytes.Length);
            // Save the OCR result as a text file
            fs.Close(); 
        }
        MessageBox.Show("OCR successful");
    } 
    else 
    {
        MessageBox.Show(dynamicDotNetTwain1.ErrorString); 
    }
}

private void button1_Click(object sender, EventArgs e) 
{
    OpenFileDialog filedlg = new OpenFileDialog(); 
    if (filedlg.ShowDialog() == DialogResult.OK) 
    { 
        dynamicDotNetTwain1.LoadImage(filedlg.FileName);
        // Choose an image from your local disk and load it into Dynamic .NET TWAIN
    } 
}

private void dynamicDotNetTwain1_OnImageAreaSelected(short sImageIndex, int left, int top, int right, int bottom) 
{
    dynamicDotNetTwain1.OCRTessDataPath = "../../"; 
    dynamicDotNetTwain1.OCRLanguage = "eng";
    OcrResultFormat ocrResultFormat = Dynamsoft.DotNet.TWAIN.OCR.ResultFormat.Text;

    byte [] sbytes = dynamicDotNetTwain1.OCR(dynamicDotNetTwain1.CurrentImageIndexInBuffer, left, top, right, bottom);
    // OCR the selected area of the image

    if (sbytes != null) 
    { 
        SaveFileDialog filedlg = new SaveFileDialog(); 
        filedlg.Filter = "Text File(*.txt)| *.txt"; 
        if (filedlg.ShowDialog() == DialogResult.OK) 
        {
            FileStream fs = File.OpenWrite(filedlg.FileName); 
            fs.Write(sbytes, 0, sbytes.Length);
            // Save the OCR result as a text file
            fs.Close(); 
        }
        MessageBox.Show("OCR successful");
    } 
    else 
    {
        MessageBox.Show(dynamicDotNetTwain1.ErrorString); 
    }
}

Private Sub button1_Click(ByVal sender As Object, ByVal e As EventArgs)
	Dim filedlg As New OpenFileDialog()
	If filedlg.ShowDialog() = DialogResult.OK Then
		dynamicDotNetTwain1.LoadImage(filedlg.FileName)
		' Choose an image from your local disk and load it into Dynamic .NET TWAIN
	End If
End Sub

Private Sub dynamicDotNetTwain1_OnImageAreaSelected(ByVal sImageIndex As Short, ByVal left As Integer, ByVal top As Integer, ByVal right As Integer, ByVal bottom As Integer)
	dynamicDotNetTwain1.OCRTessDataPath = "../../"
	dynamicDotNetTwain1.OCRLanguage = "eng"
	Dim ocrResultFormat As OcrResultFormat = Dynamsoft.DotNet.TWAIN.OCR.ResultFormat.Text

	Dim sbytes() As Byte = dynamicDotNetTwain1.OCR(dynamicDotNetTwain1.CurrentImageIndexInBuffer, left, top, right, bottom)
	' OCR the selected area of the image

	If sbytes IsNot Nothing Then
		Dim filedlg As New SaveFileDialog()
		filedlg.Filter = "Text File(*.txt)| *.txt"
		If filedlg.ShowDialog() = DialogResult.OK Then
			Dim fs As FileStream = File.OpenWrite(filedlg.FileName)
			fs.Write(sbytes, 0, sbytes.Length)
			' Save the OCR result as a text file
			fs.Close()
		End If
		MessageBox.Show("OCR successful")
	Else
		MessageBox.Show(dynamicDotNetTwain1.ErrorString)
	End If
End Sub

This is how the application looks:

The quality of the input image is the most crucial determinant in the speed of an OCR task. The lower the background noise and the higher the dpi, with a great goal value of around 200 dpi, the faster and more accurate the OCR output.

Image Processing Techniques for Dynamsoft OCR

We need to use OCR in a variety of situations, such as scanning a credit card number with our phone or extracting text from paper documents. OCR capabilities are included in Dynamsoft Label Recognition (DLR) and Dynamic Web TWAIN (DWT).

Although they can do an excellent job in general, we can improve the results by using various image processing techniques.

Lighten/remove shadows

Poor illumination may have an impact on the OCR result. To improve the outcome, we can whiten photos or eliminate shadows from images.

Invert

Because the OCR engine is often trained on text in dark colors, text in light colors can be harder to discover and recognize.

It will be easier to recognize if we invert its color

To perform the inversion, we can use the GrayscaleTransformationModes parameter in DLR.

Here are the JSON settings:

"GrayscaleTransformationModes": [
    {
        "Mode": "DLR_GTM_INVERTED"
    }
]

DLR .net’s reading result:

Rescale

If the letter height is too low, the OCR engine may not produce a good result. In general, the image should have a DPI of at least 300.

There is a ScaleUpModes parameter in DLR 1.1 that allows you to scale up letters. We may, of course, scale the image ourselves.

Reading the image directly yields the incorrect result:

After scaling up the image x2, the result is correct:

Deskew

It is fine if the text is a little distorted. However, if it is overly skewed, the outcome will be adversely altered. To improve the outcome, we need to crop the image.

To accomplish this, we can use the Hough Line Transform in OpenCV and Python.

Here is the code to deskew the image above:

import cv2
import numpy as np
import math
from PIL import Image

def deskew():
    src = cv2.imread("neg.jpg", cv2.IMREAD_COLOR)
    gray = cv2.cvtColor(src, cv2.COLOR_BGR2GRAY)
    kernel = np.ones((5,5), np.uint8)
    erode_img = cv2.erode(gray, kernel)
    eroDil = cv2.dilate(erode_img, kernel)
    show_and_wait_key("eroDil", eroDil)

    canny = cv2.Canny(eroDil, 50, 150)
    show_and_wait_key("canny", canny)

    lines = cv2.HoughLinesP(canny, 0.8, np.pi / 180, 90, minLineLength=100, maxLineGap=10)
    drawing = np.zeros(src.shape[:], dtype=np.uint8)

    maxY = 0
    degree_of_bottomline = 0
    index = 0
    for line in lines:
        x1, y1, x2, y2 = line[0]
        cv2.line(drawing, (x1, y1), (x2, y2), (0, 255, 0), 1, lineType=cv2.LINE_AA)
        k = float(y1-y2)/(x1-x2)
        degree = np.degrees(math.atan(k))
        if index == 0:
            maxY = y1
            degree_of_bottomline = degree
        else:        
            if y1 > maxY:
                maxY = y1
                degree_of_bottomline = degree
        index += 1
    show_and_wait_key("houghP", drawing)

    img = Image.fromarray(src)
    rotate_img = img.rotate(degree_of_bottomline)
    rotate_img_cv = np.array(rotate_img)
    cv2.imshow("rotateImg", rotate_img_cv)
    cv2.imwrite("deskewed.jpg", rotate_img_cv)
    cv2.waitKey()

def show_and_wait_key(win_name, img):
    cv2.imshow(win_name, img)
    cv2.waitKey()

if __name__ == "__main__":
    deskew()

import cv2
import numpy as np
import math
from PIL import Image

def deskew():
    src = cv2.imread("neg.jpg", cv2.IMREAD_COLOR)
    gray = cv2.cvtColor(src, cv2.COLOR_BGR2GRAY)
    kernel = np.ones((5,5), np.uint8)
    erode_img = cv2.erode(gray, kernel)
    eroDil = cv2.dilate(erode_img, kernel)
    show_and_wait_key("eroDil", eroDil)

    canny = cv2.Canny(eroDil, 50, 150)
    show_and_wait_key("canny", canny)

    lines = cv2.HoughLinesP(canny, 0.8, np.pi / 180, 90, minLineLength=100, maxLineGap=10)
    drawing = np.zeros(src.shape[:], dtype=np.uint8)

    maxY = 0
    degree_of_bottomline = 0
    index = 0
    for line in lines:
        x1, y1, x2, y2 = line[0]
        cv2.line(drawing, (x1, y1), (x2, y2), (0, 255, 0), 1, lineType=cv2.LINE_AA)
        k = float(y1-y2)/(x1-x2)
        degree = np.degrees(math.atan(k))
        if index == 0:
            maxY = y1
            degree_of_bottomline = degree
        else:        
            if y1 > maxY:
                maxY = y1
                degree_of_bottomline = degree
        index += 1
    show_and_wait_key("houghP", drawing)

    img = Image.fromarray(src)
    rotate_img = img.rotate(degree_of_bottomline)
    rotate_img_cv = np.array(rotate_img)
    cv2.imshow("rotateImg", rotate_img_cv)
    cv2.imwrite("deskewed.jpg", rotate_img_cv)
    cv2.waitKey()

def show_and_wait_key(win_name, img):
    cv2.imshow(win_name, img)
    cv2.waitKey()

if __name__ == "__main__":
    deskew()

Lines detected:

Deskewed:

Image Processing Techniques for IronOCR

The quality of the input image is not important here because IronOCR excels at repairing defective documents (though this is time-consuming and will cause your OCR jobs to use more CPU cycles).

Choosing input image formats with less digital noise, such as TIFF or PNG, can also result in speedier outcomes than lossy image formats, such as JPEG.

The image filters listed below can significantly enhance performance:

• OcrInput.Rotate(double degrees) — Rotates images clockwise by a specified number of degrees. Negative integers are used for anti-clockwise rotation.
• OcrInput.Binarize() — This image filter makes every pixel either black or white, with no in-between. It may improve OCR performance in circumstances where the text-to-background contrast is very low.
• OcrInput.ToGrayScale() — This image filter converts every pixel to a grayscale shade. It is unlikely to improve OCR accuracy, but it may increase speed.
• OcrInput.Contrast() — Automatically increases contrast. In low-contrast scans, this filter frequently improves OCR speed and accuracy.
• OcrInput.DeNoise() — This filter should be used only when noise is expected.
• OcrInput.Invert() — Reverses all colors. For example, white becomes black: black becomes white.
• OcrInput.Dilate() — Advanced morphology. Dilation is the process of adding pixels to the edges of objects in an image. (Inverse of Erode)
• OcrInput.Erode() — An advanced morphology function. Erosion is the process of removing pixels from the edges of objects. (Inverse of Dilate)
• OcrInput.Deskew() — Rotates an image so that it is orthogonal and the right way up. Because Tesseract tolerance for skewed scans can be as low as 5 degrees, this is quite useful for OCR.
• DeepCleanBackgroundNoise() — Removes a lot of background noise. Only use this filter if you know there is a lot of background noise in the document because it can reduce OCR accuracy on clear documents and is quite CPU intensive.
• OcrInput.EnhanceResolution — Improves the resolution of low-resolution photos. Because of OcrInput, this filter is rarely used. OcrInput will detect and resolve low resolution automatically.

We may want to use IronTesseract to speed up OCR on higher-quality scans. If we're looking for speed, we might start here and subsequently turn features back on until the proper balance is struck.

using IronOcr;

class Program
{
    static void Main(string[] args)
    {
        var ocr = new IronTesseract();

// Configuration for speed tuning
        ocr.Configuration.BlackListCharacters = "~`$#^*_}{][|\\";
        ocr.Configuration.PageSegmentationMode = TesseractPageSegmentationMode.Auto;
        ocr.Configuration.TesseractVersion = TesseractVersion.Tesseract5;
        ocr.Configuration.EngineMode = TesseractEngineMode.LstmOnly;
        ocr.Language = OcrLanguage.EnglishFast;

        using (var input = new OcrInput(@"img\Potter.tiff"))
        {
            var result = ocr.Read(input);
            Console.WriteLine(result.Text);
        }
    }
}

using IronOcr;

class Program
{
    static void Main(string[] args)
    {
        var ocr = new IronTesseract();

// Configuration for speed tuning
        ocr.Configuration.BlackListCharacters = "~`$#^*_}{][|\\";
        ocr.Configuration.PageSegmentationMode = TesseractPageSegmentationMode.Auto;
        ocr.Configuration.TesseractVersion = TesseractVersion.Tesseract5;
        ocr.Configuration.EngineMode = TesseractEngineMode.LstmOnly;
        ocr.Language = OcrLanguage.EnglishFast;

        using (var input = new OcrInput(@"img\Potter.tiff"))
        {
            var result = ocr.Read(input);
            Console.WriteLine(result.Text);
        }
    }
}

Imports IronOcr

Friend Class Program
	Shared Sub Main(ByVal args() As String)
		Dim ocr = New IronTesseract()

' Configuration for speed tuning
		ocr.Configuration.BlackListCharacters = "~`$#^*_}{][|\"
		ocr.Configuration.PageSegmentationMode = TesseractPageSegmentationMode.Auto
		ocr.Configuration.TesseractVersion = TesseractVersion.Tesseract5
		ocr.Configuration.EngineMode = TesseractEngineMode.LstmOnly
		ocr.Language = OcrLanguage.EnglishFast

		Using input = New OcrInput("img\Potter.tiff")
			Dim result = ocr.Read(input)
			Console.WriteLine(result.Text)
		End Using
	End Sub
End Class

This result is 99.8% accurate compared to the baseline of 100% — but 35% faster.